22/07/2026 10:25 pm
Topic starter
Notifications
Clear all
DeepSeek Forum
3
Posts
4
Users
0
Reactions
59
Views
0
Ive been running local LLMs for ages but DeepSeek V4 Flash is giving me a headache. Trying to fit it on my 3090 rig before my project deadline next week and standard GGUF is just tanking performance. Has anyone tried EXL2 versus AWQ for this specifically? Really dont want to sacrifice too much perplexity here...
3 Answers
12
22/07/2026 11:30 pm
If you are strictly racing against a deadline, stick with AutoAWQ 4-bit quantization. It is way easier on your VRAM overhead than EXL2, which gets super picky with context window sizing. You might want to be careful about pushing the context too far, or you will hit an OOM error instantly. Basically, AWQ is just the safest bet for that NVIDIA GeForce RTX 3090 24GB GDDR6X setup.
10
22/07/2026 10:25 pm
Honestly, I had issues with DeepSeek on my NVIDIA GeForce RTX 3090 24GB GDDR6X. Using GGUF was a total disaster for me too.
- EXL2 feels faster but unstable, crashed on me twice.
- AWQ is way more reliable for my workflow. I just stick to AWQ now, even if it feels a bit slower. Its just not worth the headache of constant errors when youre on a deadline. Hope you get it running.
3
22/07/2026 10:40 pm
To add to the point above: AWQ is definitely the way to go! Just watch your VRAM usage, cause hitting that memory limit is a nightmare for stability. Keep it chill!
